Two line graphs compare long-term degradation prediction models. The first graph shows the long-term degradation prediction behavior over time, with the x-axis representing time in days and the y-axis representing volume-averaged property retention. The second graph zooms in on the short-term training interval from 0 to 180 days. The models compared include PINN, ANN, SVR, and linear regression. The PINN model shows smoother and more physically interpretable degradation trajectories compared to the other models. All values are approximated.Comparison of long-term degradation prediction behavior between the proposed physics-informed neural network (PINN) framework and baseline data-driven models at 453 K. Figure 2(a) illustrates the long-horizon extrapolation behavior over the extended prediction domain, where the vertical dashed line at 180 days denotes the boundary between the short-term training window and the extrapolation regime. Although all models exhibit relatively similar interpolation behavior within the observed domain, purely data-driven models display increased extrapolation sensitivity, oscillatory behavior, or reduced physical consistency during long-term prediction. The proposed PINN maintains comparatively smooth and physically interpretable degradation trajectories because the governing Arrhenius and reaction–diffusion physics are explicitly enforced during training. Figure 2(b) presents a zoomed comparison within the short-term training interval (0–180 days), demonstrating that all models achieve relatively comparable interpolation performance during the observed aging period, thereby explaining the similar RMSE values obtained within the short-term test domain despite substantial divergence during long-horizon extrapolation